ModelBrief · Gemini 4 Neon · Topic Page
Last updated Nov 20, 2026
USA · LLM · GA · NOV 18, 2026

Gemini 4 Neon: The Topic Page

Google's mid-tier Gemini 4 model, released November 18, 2026: it sits above the fast Hydrogen tier and below the flagship Carbon, with the official pitch of "Carbon-level quality at Flash speeds". It scores 61 on the Artificial Analysis Intelligence Index (xhigh effort) for third place, leads the AA Coding Agent Index at 58.6, and costs $3/$15 per 1M tokens with a 1M context and four thinking levels. Two things first, though: the 61 needs the most expensive xhigh tier — at the default high tier it scores 60, the same as Carbon — and until November 30, Neon costs 50% more per token than Carbon's promotional price; "cheaper than Carbon" only becomes true per token on December 1. This page collects the official specs, the ways to try it, pricing, and the first coverage — with unverified claims clearly marked.

61AA Intelligence Index v5.0.1 (xhigh, #3)
$3 / $15API per 1M tokens (input / output)
1M tokensContext window (1,048,576)
65,536Max output tokens (sync default; long decode up to 1M)
low / med / high / xhighThinking levels (high default)
No free tierPaid API or Pro / Ultra subscription

01Key facts

What is confirmed: this is a general release of a mid-tier model, not a preview — but its headline score of 61 needs the most expensive thinking tier.

Announced / released
Officially released 2026-11-18 (Koray Kavukcuoglu, SVP Google DeepMind / Google Chief AI Architect, blog.google); live the same day, no waitlist
Model ID
gemini-4-neon (stable API identifier)
Positioning
Mid-tier of the Gemini 4 family: "frontier-level intelligence with fast end-to-end task completion" for complex software engineering, knowledge work and science; Carbon takes the hardest long-horizon projects, Hydrogen handles fast everyday execution, Neon balances depth and efficiency
Built on
The model card states it is adapted from Gemini 4 Carbon; architecture, training data and hardware details all point to the Carbon model card
Context
1,048,576 tokens, unified input + output window
Max output
65,536 tokens (synchronous default); up to 1,048,576 tokens with long decode
Reasoning
Four levels: low / medium / high (default) / xhigh; Carbon has three — Neon adds xhigh
Input / output
Input: text, image, audio, video (plus PDF). Output: text. No image generation, audio generation or Live API
Knowledge cutoff
March 2026 (per the model card)
Access
Gemini API paid tier (AI Studio, Android Studio), Gemini for business (API and model dropdown), Google Antigravity (paid Pro / Ultra subscribers), Gemini app (Pro / Ultra subscribers). No free tier; Gemini 4 Neon for cyber rolls out to trusted Fairwind Program defenders in the coming days
Price
API: $3 in / $15 out per 1M tokens, $0.15 cached input; Batch / Flex $1.50 / $7.50; Priority $5.40 / $27.00. No promotional price — the launch-day price is the fixed price
Benchmarks
Artificial Analysis index 61 (xhigh, v5.0.1, #3); AA Coding Agent Index 58.6, #1; Vals Index 70.9%, #2; Agent Arena Index 1336, #2; LMArena Text 1518, #3 (snapshot, see section 12); the model card's 18-test table leads in most tests (all Google's own runs, see section 12)
Weights / license
None published; proprietary hosted model, not open source
Parameters
Not disclosed

02Quick comparison

Where it stands among its closest peers, at a glance.

ModelAA IndexPrice (input · output, per 1M tokens)ContextOpen source
Gemini 4 Neon61 (xhigh)$3 / $151M tokensNo
Claude Fable 5.265 (Max tier)$8 / $401M tokensNo
GPT-6.2 Astra62 (Max tier)$8 / $401M tokensNo
Gemini 4 Carbon60 (High tier)$2 / $10 (until Nov 30), $4 / $20 from Dec 11M tokensNo
Gemini 4 Hydrogen54 (xhigh tier)$1.50 / $7.50 (has a free tier)1M tokensNo
Kimi K3.1 Pro46$2 / $101M tokensYes (custom license)

AA Index is Artificial Analysis' Intelligence Index v5.0.1, each model at its highest thinking tier; Neon scores 60 at its default high tier — the same as Carbon — and 61 only at xhigh. Prices checked November 2026 — the checkout page is authoritative. For the full point-by-point comparison, see the Comparisons section.

03Charts

The independent index first; Google's own benchmark table follows in the Benchmarks section.

Bar chart, AA Intelligence Index: Claude Fable 5.2 65, GPT-6.2 Astra 62, Gemini 4 Neon 61, Gemini 4 Carbon 60, Gemini 4 Hydrogen 54, Kimi K3.1 Pro 46
Artificial Analysis Intelligence Index v5.0.1: Gemini 4 Neon (orange) at 61, third — one point ahead of Carbon (60), one behind Astra (62). Source: Artificial Analysis.

04Top 3 reads

Short on time? Start with these three.

05Timeline

Two days from rumor to announcement; the name "Neon" had already appeared once before, in May — for something else entirely.

  • 2026-05-05

    "Neon" first appears: 9to5Google's APK Insight team finds traces of a subscription plan codenamed Neon in a macOS Gemini app update, tentatively called Google AI Ultra Lite (a tier between Pro and Ultra). That was a subscription codename, not a model — Google later reused the name for this model.

  • 2026-11-16

    TestingCatalog finds a reference to gemini-4-neon in Gemini web-app code; the model is not yet available. The author explicitly labels "new Flash Lite tier" as pure speculation. Chinese forums and tracker accounts repeat the finding without adding sources.

  • 2026-11-18, early hours

    Google Cloud's billing catalog leaks Neon's pricing ahead of the announcement: $3 in / $15 out per 1M tokens, $0.15 cached input, effective 3 AM ET; Antigravity CLI 1.5.11 adds a gemini-pro-orchestrator-neon orchestrator variant the same day.

  • 2026-11-18

    Official launch: Google announces Gemini 4 Neon on blog.google (by Koray Kavukcuoglu); the model card goes live the same day; AA index 61 (xhigh) for third place, Coding Agent Index 58.6 at No. 1; the API, AI Studio, Antigravity and the Gemini app all ship the same day. The November 16 rumor status ends here.

  • 2026-11-18

    First English coverage: VentureBeat, 9to5Google and The Decoder publish the same day; 9to5Google reports Neon replaces Argon in the Gemini app and 3.1 Pro is going away — note that Google's own blog does not say either, so this is attributed to 9to5Google (see section 15).

  • 2026-11-18

    OpenRouter lists Neon: model ID google/gemini-4-neon, at the same prices as the official API — $3/$15 per 1M tokens, $0.15 cached input — with a 1,048,576-token context; providers are Google AI Studio and Vertex (see section 07).

  • 2026-11-18

    CursorBench 4.1 (Cursor changelog): Neon leads at Extra High with 65.4% at $10.67 per task, and is #2 at High with 63.7% at $6.54 per task — matching the 65.4% Google itself reports in the model card (see section 12).

  • 2026-11-19

    Read by this site: Artificial Analysis' model page confirms the four-tier scores (xhigh 61 / high 60 / medium 55 / low 44) and timing data; the LMArena Text snapshot is 1518, third (3,885 votes).

  • 2026-11-18

    Artificial Analysis breaks down Neon's index: on AA-Briefcase (agentic knowledge work) Neon scores 1,700 Elo overall, and its Presentation Quality of 1,771 Elo is the highest Artificial Analysis has measured, 45 points ahead of the previous leader, Sonnet 5.5 (1,726); on GDP.pdf the high tier scores 33.4%, ahead of Carbon's 28.4% (xhigh 28.0%). Overall Briefcase Elo still trails Fable 5.2 (1,785) and Sonnet 5.5 (1,815). See section 12.

  • 2026-11-19/20

    KingBench 4 (Peter Gostev, YouTube video): six one-prompt app builds; Neon scores 37.9/60 (63.2%), #2 of 18 models, behind only Fable 5.2 (42.0); one tester, a six-task mini-suite — see section 09.

  • 2026-11-30

    Introductory API pricing for Carbon and Argon ends ($2/$10).

  • 2026-12-01

    Carbon and Argon API prices rise to $4 input / $20 output per 1M tokens; from this day, Neon is 25% cheaper per token than Carbon.

  • Coming days

    Gemini 4 Neon for cyber rolls out to Fairwind Program trusted defenders; the full Neon Safety & Security Report is due to be published.

06Official specs

Official figures from the model page and model card (checked Nov 18–19, 2026); unpublished items are stated as such, never filled in by guesswork.

Status
General release (since 2026-11-18, no preview label)
Model ID
gemini-4-neon
Context window
1,048,576 tokens (unified input + output window)
Max output
65,536 tokens (synchronous default); up to 1,048,576 tokens with long decode
Reasoning
Four levels: low / medium / high (default) / xhigh; Carbon has three
Input
Text, image, audio, video (plus PDF)
Output
Text (no image generation, audio generation or Live API)
Knowledge cutoff
March 2026
API
Gemini API paid tier; no free tier
Benchmarks
AA index 61 (xhigh, v5.0.1); AA Coding Agent Index 58.6, #1; LMArena Text 1518 (snapshot)
Model card
Published (2026-11-18); Safety & Security Report pending
Parameters
Not disclosed
Weights
Not published
API price
$3 / $15 per 1M tokens, no promotional price; $0.15 cached input

07How to try it: 6 routes

Developers: the paid API is fastest. Everyone else: a Pro subscription is the minimum — free users can't reach Neon; Hydrogen is the one with a free tier.

1Gemini API paid tier (AI Studio / Android Studio)

The main developer route. Any paid-tier API key can call it as gemini-4-neon; migrating from another Gemini 4 model means changing only the model string. There is no free tier — you need a billing account with prepaid credit.

2Gemini app (Pro / Ultra subscription)

Google AI Pro and Ultra subscribers pick it from the model dropdown in the Gemini app. 9to5Google reports the refreshed picker has four entries: Auto, Everyday help (4 Hydrogen), Deeper analysis (4 Neon), Code & complex work (4 Carbon), with Neon taking Argon's place. Google's official blog does not mention the picker change — see section 15 for the attribution.

3Google Antigravity (paid Pro / Ultra subscription)

Available in the desktop app, CLI and IDE extensions; Neon can act as an Orchestrator directing parallel Hydrogen subagents on long coding jobs. Antigravity CLI 1.5.11 added the gemini-pro-orchestrator-neon variant.

4Gemini for business

The enterprise route: API or model dropdown. Trackers report Neon is already listed in the Vertex AI Model Garden and Google Cloud's billing catalog; Google's official blog only says "Gemini for business" and does not name Vertex AI separately — see section 14.

5Neon for cyber (Fairwind Program)

The cybersecurity variant: rolling out to trusted Fairwind Program defenders in the coming days, by application through a Cloud account team. It targets secure code generation, penetration testing, vulnerability finding and patching; SEC-bench Pro 90.5% and AutoBaxBench hard tier 76% (both Google-reported, see section 12).

6OpenRouter

A third-party aggregation gateway: model ID google/gemini-4-neon, at the same prices as the official API — $3 in / $15 out per 1M tokens, $0.15 cached input — with a 1,048,576-token context; providers are Google AI Studio and Vertex. OpenRouter measures a median throughput of 88 tok/s (P50).

Model: gemini-4-neon thinking_level: low | medium | high (default) | xhigh max_output_tokens: 65,536 default synchronous; long decode up to 1,048,576 HTTP timeout: allow plenty (AA measured 21.5 min per index task at xhigh, ~56.5s to first answer token)
Getting-started warning (AA measurements): at xhigh, each index task averages 21.5 minutes with about 56.5 seconds to the first answer token — far slower than the same-price-tier medians (12.4 minutes, 13.08 seconds). "Fast end-to-end completion" is about task cost relative to Carbon, not absolute speed. Also note: like Carbon, Neon ignores temperature, topP and topK, and supports only implicit caching (no explicit caching).

08Pricing

The price list itself is simple; the calendar is what makes it complicated: Neon's $3/$15 is a fixed price, while Carbon's $2/$10 is a promotional price that ends November 30. Prices checked Nov 18–19, 2026 — the checkout page is authoritative.

API tier (per 1M tokens)InputOutput (incl. thinking tokens)Cached input
Standard$3.00$15.00$0.15
Batch API$1.50$7.50$0.075
Flex$1.50$7.50$0.075
Priority$5.40$27.00$0.27

Caching is implicit: it triggers automatically on prompts of 4,096 tokens or more; explicit caching is not supported. Batch and Flex are half the standard price; Priority is 1.8x standard.

The calendar arithmetic: three sums

First, Neon costs exactly 2x Hydrogen ($3/$15 vs $1.50/$7.50), and Hydrogen has a free tier while Neon has none. Second, until November 30, Neon costs 50% more per token than Carbon ($3/$15 vs the promotional $2/$10) — so "cheaper than the flagship" is not true per token on launch day. Third, from December 1, Carbon and Argon rise to $4/$20 and Neon becomes 25% cheaper than them. The same price list gives opposite conclusions in November and December.

Cost per task: Google's claim vs another snapshot

Google says Neon costs 32% less per task than Carbon at the default high tier, and 36%+ less than top competing models at their highest settings — Google's own measurement on AA index tasks. But LMArena's Agent cost snapshot (via kingy.ai, Nov 18) points the other way: a median $2.66 per task for Neon against $1.34 for Carbon. The two use different methods (AA index tasks vs Arena live sessions) and cannot be reconciled yet, so this page treats neither as settled. See our pricing guide.

Pricing takeaway: to be cheaper than Carbon per token, wait until December 1; to use it now, remember there is no promotional price and no free tier — Batch / Flex at half price is the only certain discount. Control cost with the thinking tier: high uses less than half the output tokens of xhigh on the AA index (123M vs 283M) and loses only 1 point.

Official pages: Gemini API pricing · Google's Nov 18 launch post

09First coverage and hands-on tests

Two days old: mostly launch coverage and analysis from Nov 18–19, plus the first independent hands-on test, KingBench 4, from Nov 19–20. Treat the coverage as reporting; the test is a small sample — read it as a case study. Each item carries a verdict badge; "read our full article" links go to this site's in-depth articles, with the original source at the foot of each article.

English sources

VentureBeat: Opus 5.5-level coding at 25% lower token prices

Verdict: launch report, not a test

Source: VentureBeat, Nov 18, 2026 (launch report)

Frames Neon's pitch as "Opus 5.5-level coding performance at 25% lower API token prices"; notes that Google's launch post gives no absolute completion times in seconds or minutes for a matched Neon-vs-Carbon workload, so "faster" is qualitative for now. Also cites Artificial Analysis confirming on X that Neon placed third on the Intelligence Index.

Read the original → (venturebeat.com)

9to5Google: Carbon quality, Flash speeds, for AI Pro

Verdict: launch report, with exclusive picker details

Source: 9to5Google, Nov 18, 2026 (launch report)

The headline is the positioning: "Carbon quality at Flash speeds". The exclusive detail is the Gemini app's new picker: Neon replaces Argon and 3.1 Pro is going away, leaving Auto / Everyday help (4 Hydrogen) / Deeper analysis (4 Neon) / Code & complex work (4 Carbon). It also confirms four thinking levels (Low / Medium / High / Extra High) and the $3/$15 price — twice Hydrogen's.

Read the original → (9to5google.com)

The Decoder: the fourth Gemini 4 model in a month

Verdict: launch report, not a test

Source: The Decoder, Nov 18, 2026 (launch report)

Stresses the pace: Neon is the fourth Gemini 4 model in four weeks. Confirms the 1M context, long output up to 1M tokens, Computer Use and code execution support, and no image or audio generation. Benchmark figures had not been published when it filed (Google released the full tables in the blog and model card later that evening).

Read the repost → (zerohour.day repost)

kingy.ai: line-by-line model card check + LMArena snapshot

Verdict: analysis, not a test

Source: kingy.ai, Nov 18, 2026 (model card and leaderboard analysis)

Checks DeepMind's model card table line by line (Neon vs Opus 5.5 and GPT-6.2 Sol) and captures LMArena snapshots: Neon High at 1518 on Text, third (3,885 votes, rank range 1–5), WebDev at 1,764, #6 (2,308 votes), Agent Arena on 4,234 sessions (rank range 2–6). Also notes the model card calls Neon "adapted from Carbon" with a March 2026 knowledge cutoff.

Read the original → (kingy.ai)

TestingCatalog / AI Frontier Post: the full rumor-to-launch timeline

Verdict: rumor recap, not a test

Source: TestingCatalog, AI Frontier Post, Nov 16–18, 2026

Found the gemini-4-neon string in Gemini web code on November 16 and explicitly labelled "Flash Lite" as pure speculation; the $3/$15 price leaked in the billing catalog in the early hours of the 18th; the evening recap admits the story flipped from rumor to launch within a day — and notes there is no evidence for the "Krypton renamed" theory.

Read the original → (aifrontierpost.com)

KingBench 4: six one-prompt app builds, Neon 37.9/60, #2 of 18

Verdict: the first independent hands-on test puts it #2, behind only Fable 5.2

Source: Peter Gostev (YouTube video), Nov 19–20, 2026 (one tester, a six-task mini-suite; 18 models in the same run)

All six tasks are "one prompt, build a working app from scratch": a Rust simulator, a Go container runtime, a 3D room designer, a QMK keyboard configurator, a Pokémon Red remake, and a Rust Fitbit desktop app. Neon scores 37.9/60 (63.2%), #2 of 18 models, behind only Fable 5.2 (42.0) and ahead of Opus 5.5 (34.7), Astra (33.6), Carbon (32.0) and Hydrogen (24.9); the whole round costs $20.36. The six sub-scores, in order, are 4.3, 3.0, 8.5, 6.2, 8.4 and 7.5. The limits must be stated: a single tester and six tasks, a small sample, in video form with scoring details not fully disclosed — not a verdict.

Watch the original video → (youtube.com, video)

Chinese sources

Agentic Commons: almost every number in the announcement is Google's own measurement

Verdict: analysis, flags the fine print

Source: Agentic Commons, Nov 18, 2026 (Traditional Chinese analysis)

After checking the official announcement line by line, it warns that almost all performance and cost figures come from Google's own measurements: the same Neon scores 12.8% on SWE-sweep with the third-party mini-SWE harness but 39.0% inside Google's Antigravity harness — the harness matters enormously. It also noted that the DeepMind model card page showed only a title at the time of its check (the card was complete when this site read it on Nov 19).

Read the original → (agenticcommons.xyz, Traditional Chinese)

Marketersgo: headline claims "3x faster, 40% cheaper"

Verdict: headline overstates; do not cite it

Source: Marketersgo (商傳媒), Nov 18, 2026 (Traditional Chinese report)

The body matches the official announcement (AA 61 third, Vals 70.9% second, $3/$15), but the headline "3x faster and 40% cheaper" has no basis in Google's announcement: Google claims only a 32% lower cost per task than Carbon at the default high tier, and never published a "3x speed" figure. This site does not use the headline claim — see section 15.

Read the original → (marketersgo.com, Traditional Chinese)

10Articles: our in-depth coverage

The 6 articles below are this site's original rewrites and reorganizations — restructured, with conclusions and buying advice added, and facts and figures checked against the original sources. Click any article to read it on this site; the original source is cited at the foot of each article.

Gemini 4 Neon Launch: From Rumor to Nov 18 ReleaseOur reconstruction of the launch: TestingCatalog finds the model name in Gemini web code on November 16, the billing catalog leaks the price in the early hours of the 18th, and Google announces it officially that evening and ships the same day — plus who actually reported the Gemini app picker change. Gemini 4 Neon Benchmarks, Fully Mapped: AA 61 Across Four Tiers, the 18-Test Model Card Table, and the Independent IndexesOur roundup of every verifiable score: AA index xhigh 61 / high 60 / medium 55 / low 44, Coding Agent Index 58.6 at No. 1, Vals and Agent Arena at No. 2 — and which of the model card's 18 tests are Google's own runs, and which ones Neon loses to Opus 5.5. Gemini 4 Neon Pricing: What $3/$15 Means Before and After December 1Our three sums: Neon costs 2x Hydrogen, 50% more than Carbon until November 30, and 25% less than Carbon from December 1 — plus why Google's "32% cheaper per task" claim and LMArena's cost snapshot point in opposite directions. Gemini 4 Neon vs Carbon: Same Lineage, 61 vs 60, and When Switching to Neon PaysOur point-by-point comparison: Neon is adapted from Carbon, leads the AA index by 1 point only at xhigh, ties at the default tier, and the pricing calendar decides which is cheaper; with a decision table. Gemini 4 Neon vs Hydrogen: What 2x the Price BuysOur point-by-point comparison: AA index 61 vs 54, four thinking tiers each, the free-tier gap, speed and task cost; with a decision table. Gemini 4 Neon Access and API Guide: Model ID, Thinking Tiers, Long Decode and Caching RulesOur map of every access route: the API model ID, AI Studio, Antigravity, the Gemini app picker, the Vertex AI attribution note, plus the long-decode, implicit-caching and ignored-parameter pitfalls.

12Benchmarks: independent index + vendor-reported

Neon's scores come in two kinds: AA, Vals and Agent Arena are independent indexes; the model card's 18-test table is Google's own runs. Independent first, then Google's table, then the caveats.

Artificial Analysis Intelligence Index: four tiers

Thinking tierAA Index v5.0.1Time per index taskTime to first answer tokenOutput tokens per index run
xhigh (top tier)61, #321.5 min56.51s283M
high (default)6014.8 min40.64s123M
medium558.3 min20.14s62.9M
low443.7 min13.93s28.3M

Source: Artificial Analysis model page (read Nov 19, 2026). The AA index combines 9 evaluations (AA-Briefcase, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode-AA 2.0, AA-Omniscience, GDP.pdf, HLE-Diamond, CritPt v2, Terminal-Bench-Science 0.1). xhigh beats high by 1 point while generating 2.3x the output tokens and taking 45% longer.

Independent indexes and third-party measurements

MeasurementSourceNeon resultComparison
AA Intelligence Index v5.0.1Artificial Analysis, Nov 1861 (xhigh), #3Fable 5.2 65, #1; Astra 62, #2; Carbon 60; Hydrogen 54
AA Coding Agent IndexArtificial Analysis58.6, #1Google's blog figure; the index's composition was not separately verified against AA's page in this round
Vals Index v2.2Vals70.9%, #2, on the Pareto frontierLegal Research Bench #1; Tax Agent Bench #1; Opus 5.5 70.1% (model card figure)
Agent Arena IndexLMArena / Agent Arena1336, #2 (4,200+ real agent sessions)Steerability #2; rank range 2–6 (kingy.ai snapshot)
LMArena TextLMArena (via kingy.ai), Nov 18 snapshot1518 ±10, #3 (3,885 votes)Carbon High 1526; Argon High 1521; rank range 1–5
LMArena WebDevLMArena (via BenchLeader), Nov 18, 2026 result1,764, #6 (high effort)2,308 votes; Fable 5.2 (xhigh) 1,824 #1, Astra (max) 1,781 #3
Surge AI (finance / accounting / spreadsheets)Surge AI (as summarized by Google's blog)#1 on most benchmarksGoogle's summary; the original boards were not read directly in this round
AA-Briefcase (agentic knowledge work)Artificial Analysis, Nov 181,700 Elo overall; Presentation Quality 1,771 Elo, the highest AA has measuredPresentation Quality leads the previous leader Sonnet 5.5 (1,726) by 45 points; overall Elo still trails Fable 5.2 (1,785) and Sonnet 5.5 (1,815); GDP.pdf high tier 33.4% beats Carbon's 28.4%
CursorBench 4.1 (Extra High)Cursor (Cursor changelog, Nov 18, 2026)65.4%, #1, $10.67 per taskMatches the 65.4% Google itself reports in the model card, but from a different source: this is the CursorBench 4.1 leaderboard result
CursorBench 4.1 (High)Cursor (Cursor changelog, Nov 18, 2026)63.7%, #2, $6.54 per task—

Google model card table (Google-reported, vs Claude Opus 5.5 and GPT-6.2 Sol)

Test (model card, 2026-11-18)NeonOpus 5.5GPT-6.2 Sol
AA Index v5.0.1615957
Vals Index v2.270.9%70.1%66.1%
CursorBench 4.165.4%58.1%—
Terminal-Bench 4.1 (cheating-adjusted)65.2%65.8%61.5%
FrontierCode 2.0 (high tier)46.2%47.3%43.8%
SWE-sweep (provider harness)39.0%10.1%8.3%
SWE-sweep (mini-SWE harness)12.8%3.6%11.6%
AutomationBench56.3%42.5%41.3%
Agents' Last Exam v1.175.8%71.6%73.0%
RiemannBench84.0%80.0%80.0%
Terminal-Bench-Science 0.170.5%63.3%68.0%
HLE-Diamond71.3%67.9%64.0%
MKA (Climb WSR)56.5%53.0%52.7%
Chartography73.5%66.3%75.1%
LVBench89.0%83.7%87.6%
GraphWalks 256K–1M (BFS F1)95.2%66.8%80.1%
CWE-Bench v270.7%67.1%66.0%
SEC-bench Pro90.5%—88.3%
WebExploitBench v1.251.3%37.1%—

Neon leads in 15 of 18 tests but loses 3: Terminal-Bench 4.1 to Opus 5.5 (65.2% vs 65.8%), FrontierCode 2.0 to Opus 5.5 (46.2% vs 47.3%), and Chartography to GPT-6.2 Sol (73.5% vs 75.1%).

Check the fine print before the scores: the model card table is Google's own runs, with rival figures taken from their reports rather than run side by side; the two SWE-sweep harnesses differ enormously (39.0% vs 12.8%), so the same model is two different stories depending on the harness — always name the harness when citing it. AA, Vals and Agent Arena are independent and more trustworthy; but the LMArena snapshot still has few votes (3,885 on Text) with a rank range of 1–5, so the score will move. LiveBench 2.0 has no Neon entry yet.

Full roundup in our benchmarks article →

13Comparisons

vs Gemini 4 Carbon (the flagship it is adapted from)

DimensionGemini 4 NeonGemini 4 Carbon
RelationshipAdapted from CarbonNeon's base model
StatusGeneral release (2026-11-18)GA (2026-11-10)
API price$3 / $15 (fixed price)$2 / $10 until Nov 30, then $4 / $20
Context1M1M
AA Index61 (xhigh) / 60 (high) / 55 / 4460 (High) / 58 (Medium) / 54 (Low)
Thinking tiersFour (adds xhigh)Three
Speed21.5 min per index task at xhighAbout 26–27 min per index task at High
SubscriptionPro / UltraUltra only
Best forHigh-frequency complex work: everyday coding agents, enterprise knowledge workThe hardest tasks: long-horizon agents, tough debugging, science, security review

Google's own framing: Carbon is for the hardest long-horizon projects; Neon is for high-frequency complex work. Both score 60 at the default tier — Neon's lead exists only at xhigh.

Read the deep comparison: Neon vs Carbon →

vs Gemini 4 Hydrogen (the fast sibling)

DimensionGemini 4 NeonGemini 4 Hydrogen
StatusGeneral release (2026-11-18)General release (2026-11-12)
API price$3 / $15$1.50 / $7.50 (half of Neon)
Free tierNoYes
Context1M1M
AA Index61 (xhigh) / 60 (high)54 (xhigh) / 51 (default tier)
Thinking tiersFour: low / medium / high / xhighFour: low / medium / high / xhigh
SubscriptionPro / UltraFree / Pro / Ultra (in the app)
Best forComplex coding, analytical reasoning, production enterprise workflowsFast everyday execution, interactive coding, quick loops

Read the deep comparison: Neon vs Hydrogen →

vs rival flagships (Artificial Analysis Index, Nov 19, 2026)

ModelAPI price $/1M (in/out)AA IndexContext
Claude Fable 5.28 / 4065 (max)1M
GPT-6.2 Astra8 / 4062 (max)1.05M
Gemini 4 Neon3 / 1561 (xhigh)1M
Gemini 4 Carbon2 / 10 (until Nov 30)60 (high)1M
Claude Opus 5.54 / 2059 (max)—
GPT-6.2 Sol2 / 1057 (max)1.05M
Gemini 4 Hydrogen1.50 / 7.5054 (xhigh)1M
Honest note: Neon's third place leads Carbon by 1 point and trails Astra by 1 — within the index's noise and tier conventions, that is effectively a tie. Its real selling points are the No. 1 Coding Agent Index and a per-token price less than half the flagships' (after December 1). Note: Opus 5.5's AA index of 59 and $4/$20 price come from Google's model card table and VentureBeat's reporting; its context window was not verified in this round, so the cell is blank.

14What's missing (as of Nov 19)

  • Parameter count and architecture details (Google discloses nothing; the model card only says "adapted from Carbon")
  • The full Neon Safety & Security Report (the model card says "in the coming days"; capability determinations currently carry forward from Argon)
  • A LiveBench 2.0 entry (no Neon score yet)
  • More independent hands-on tests (the first, KingBench 4, arrived Nov 19–20; no Quantslant- or Huxiu-style third-party review exists yet)
  • Depth coverage from mainland Chinese first-tier tech media (IT之家 / 36氪 / 虎嗅 — none found)
  • Google's own wording confirming Vertex AI (the blog only says "Gemini for business"; the Model Garden listing is tracker-reported)
  • Usage limits for Google AI Pro / Ultra subscriptions and Neon's quotas (only the eligible tiers are confirmed)
  • Longer-term pricing plans after December 1 (Google has not said whether $3/$15 is permanent)

15Rumor check

Neon had a two-day rumor window before launch, and headline writers added their own spin after it. Each claim, checked.

Resolved "Neon is just a string in Gemini's web code"

True on November 16, false from November 18: Google announced it officially that evening and shipped the API, AI Studio, Antigravity and the Gemini app the same day, with the model card and pricing page live. The rumor status is closed.

False "Neon is a new Flash Lite tier"

That was TestingCatalog's own explicitly-labelled pure speculation when it found the code string. The shipped Neon is a mid-tier model ($3/$15), twice Hydrogen's price; Hydrogen remains the cheapest tier in the family.

False "Neon is the Ultra Lite subscription"

"Neon" did appear in May 2026 as a codename for an Ultra Lite subscription tier — but that is a different thing. This launch is a model (gemini-4-neon), not a subscription tier. Google reused the codename.

False "Neon is 3x faster and 40% cheaper"

A Traditional Chinese outlet's headline. Google's announcement contains no "3x speed" figure; on cost, Google's claim is a 32% lower cost per task than Carbon at the default high tier (Google's own measurement). The headline's "3x" and "40%" should not be cited.

Partly true "Neon is cheaper than Carbon"

Per task: Google claims 32% less at the default tier — true as Google's measurement, but an LMArena cost snapshot points the other way (see section 08). Per token: false until November 30 — Neon at $3/$15 costs 50% more than Carbon's promotional $2/$10; true from December 1, when Carbon rises to $4/$20 and Neon is 25% cheaper.

False "Neon tops the AA Index"

Neon ranks third at 61, behind Fable 5.2 (65) and GPT-6.2 Astra (62). It tops the AA Coding Agent Index (58.6), not the composite index.

Unverified "Neon replaces Argon in the Gemini app, and 3.1 Pro is being retired"

This is 9to5Google's November 18 report on the picker change; Google's official blog does not mention it. The four picker entries (Auto / Everyday help / Deeper analysis / Code & complex work) likewise come only from 9to5Google. This page reports them as 9to5Google's reporting, not as Google's official wording.

Unverified "Neon is Krypton renamed"

A theory circulating online that Neon is the previously rumored Gemini 4 Krypton under a new name. No official or documentary evidence supports it; this site does not use it.

16FAQ

When was Gemini 4 Neon released?

November 18, 2026 — announced officially and live the same day. On November 16 the model name had only appeared as a string in Gemini web code (a rumor); the price leaked in the billing catalog in the early hours of the 18th, and Google announced it that evening. See the timeline in section 05.

How much does the API cost? Is there a promotional price?

Standard: $3 in / $15 out per 1M tokens, $0.15 cached input; Batch and Flex at half price ($1.50/$7.50), Priority $5.40/$27.00. There is no promotional price — the launch-day price is the fixed price. See section 08.

Is it cheaper than Carbon?

It depends: per token, no until November 30 (Carbon's promotional $2/$10 is lower); from December 1, when Carbon rises to $4/$20, Neon is 25% cheaper. Per task, Google claims 32% less at the default tier (Google's own measurement), but an LMArena cost snapshot points the other way — unsettled for now. See our pricing guide.

Is it better than Carbon?

On the AA index: Neon 61 (xhigh) vs Carbon 60 (high), a 1-point lead; at the default tier both score 60. Neon is adapted from Carbon and adds an xhigh thinking tier; it leads the Coding Agent Index at 58.6. Google's framing: Carbon is for the hardest long-horizon projects, Neon for high-frequency complex work. See our Neon vs Carbon comparison.

How is it better than Hydrogen? Which should I use day to day?

AA index 61 vs 54, stronger on complex coding and analytical reasoning — but twice the price and no free tier. Use Hydrogen for fast everyday execution and interactive coding; use Neon for high-frequency complex coding and enterprise knowledge work. See our Neon vs Hydrogen comparison.

Can free users use it?

No. The API has no free tier; the Gemini app needs a Google AI Pro or Ultra subscription; Antigravity likewise needs a paid Pro / Ultra subscription. Free users get Hydrogen in the Gemini app and the API.

How do I pick a thinking tier: low / medium / high / xhigh?

The default high tier is the best value: AA 60, 14.8 minutes per index task; xhigh scores 61 but takes 21.5 minutes and generates 2.3x the output tokens — 1 point for more than double the tokens. Medium (55, 8.3 minutes) suits mid-range work; low (44) is for the lightest jobs only. See our benchmarks article.

How big is the context window?

1,048,576 tokens, a unified input + output window. Max output is 65,536 tokens synchronously by default, and up to 1,048,576 tokens with long decode.

Is it open source? Is it on Hugging Face?

No — it is not an open model, no weights have been published, and there is no official model on Hugging Face. It is available only through Google's API and subscription services.

What happened to Argon? Is 3.1 Pro being retired?

9to5Google reports Neon replaces Argon in the Gemini app picker and 3.1 Pro is going away; Google's official blog does not mention either, so this site reports them as 9to5Google's reporting, not official wording. On the API side, Argon is still on the pricing page ($2/$10 through November 30).

Where can I follow the latest status?

Google's official blog (blog.google), the Gemini API pricing and model pages, the DeepMind model card, Artificial Analysis' model page, and the source list below.

17Sources

Every fact on this page comes from the public sources below (checked Nov 18–19, 2026).

Official

English coverage and analysis

Chinese coverage and analysis

More indexes: All Google models · All Language Models · All models